Discussion

In this work, we demonstrate how tactile feedback and a passively closing, underactuated gripper design can enable aerial perching to be more robust to target pose estimate uncertainties and target shapes and sizes. Drawing inspiration from nature, we show how in robotics the sense of touch can be a valuable complement to other senses, such as vision, especially when they are inhibited. The key contributions of this work are a compliant anthropomorphic hand equipped with soft tactile pads that act as sensorized phalanges, enabling an aerial tactile perching framework that continuously refines the MAV pose during the perching maneuver using embodied tactile feedback, as well as a tactile-based grasp validation strategy that ensures secure attachment before finalizing the perch. This work shows the importance of tactile sensing, providing feedback directly in the task space, i.e., the relative pose between the MAV and the target, as well as grasp stability, to achieve robustness in aerial robotic applications.

With the proposed tactile-based approach, the Monte–Carlo experiments show that the MAV can robustly perch with a success rate >99% for initial position offsets of up to 0.6 m, whereas the baseline only achieves ~0.1 m. This performance is also validated in experiments with the physical system, which also achieves successful perching with an initial offset up to 0.6 m. The trials with the physical prototype confirm the same offset robustness on both cylindrical and T-Bar targets, demonstrating the grasp versatility of the proposed gripper.

By adjusting its yaw angle during the perching maneuver, the MAV can also successfully align with rotational offsets of the target. From the Monte-Carlo simulation visualized in Fig. 3, we observe that both the feed-forward baseline and the proposed tactile-based approach achieve a success rate >99% for initial yaw offsets of at least up to 45°. This is mostly due to the compliance of the gripper, which naturally allows for some misalignment. However, using the tactile feedback, this bandwidth increases to ~60°. We also validate this behavior in the physical experiments, where the MAV successfully perches with an initial yaw offset of up to 20°.

Naturally, the maximum size of a perching target is constrained by the mechanical design of the gripper: longer phalanges allow the system to grasp larger targets. However, initial misalignment can prevent the gripper from fully enveloping the target, limiting its effective grasping capability. As shown in the right plot of Fig. 3, the tactile-based approach maintains a success rate above 99% for targets with radii up to 0.15 m, whereas the baseline method is limited to radii of ~0.1 m. For targets smaller than 0.02 m in radius, both methods fail to achieve successful perching. This failure is not due to insufficient enveloping, but rather because such small targets do not trigger enough contact sensors in the perched state to validate the grasp and register a successful perch.

In the perching applications outlined in this work, we expect the target to not always be aligned horizontally, but rather exhibit an inclination angle. In the Monte-Carlo simulations, we observe that both the feed-forward baseline and the proposed tactile-based approach successfully perch across the full range of inclination angles up to 45°, whereas the baseline method is slightly less reliable above 30°. This is due to the gripper’s compliance: as long as the target’s inclination remains within the gripper’s reach, each finger individually establishes contact with a different phalanx and adapts to the contact, slowly pulling the MAV underneath for a stable perch.

The tactile search pattern has a significant impact on the performance of the aerial tactile perching strategy. This process essentially constitutes a coverage path planning problem, where the MAV must systematically explore the region around the estimated target location to maximize the probability of contact. In particular, by increasing the area of the search pattern, the MAV can successfully find and perch on targets located further away from the initial estimate. However, it also increases the time required to complete the perching maneuver. As shown in the left plot of Fig. 3, success rates improve compared to the baseline method with larger initial offsets, but within the baseline method’s feasible range the perching time nearly doubles. The search pattern should therefore be matched to the expected uncertainty of the target estimate, minimizing perching time while maintaining a high success rate. Possible patterns include sinusoidal patterns, spirals, or raster scans, though platform dynamics must be considered, as patterns with sharp turns may be infeasible for the MAV. We select the sinusoidal pattern as the best compromise between success rate and perching time for the expected uncertainty of the target estimate.

The proposed tactile-based perching strategy has limitations and failure modes that need to be considered. The perching procedure relies on the assumptions A.1 and A.2. As previously discussed, too large initial offsets and target sizes violate these assumptions and will lead to failure of the searching procedure or the grasp, respectively. The same would apply for a violation of assumption A.3, where unexpected contacts would prevent the search procedure from converging to the target. In a fully integrated system, it would be expected that these assumptions are validated either by an operator via a tele-operation interface or by an onboard perception system. Another failure mode can occur at the boundary of the 95% success rate interval in positional offsets, i.e., when the target is located just on the edge of the search pattern. In these rare instances, the distal phalanx of a finger can get stuck above the target, i.e., preventing the MAV from moving underneath the target to execute the perching maneuver. However, this failure mode can be detected by monitoring the time series data of the control error in TOUCHED and APPROACH states, allowing the MAV to restart the approach from the abort state.